What 19 makers on MakerMap learned about it, in their own words: their best posts on X, filed under Building with AI.
Not every message needs Claude to think hard.
Making it red shouldn’t cost the same as building a launch video.
In my app, a cheap decision model (Cloudflare’s Clef) reads each message first and decides how hard Claude should think. It doesn’t write text, it just picks. 96% accurate on my eval
50 hours in transit, finding 2 of 10. if your coding agent pauses ten minutes between turns, the cache rules matter more than the model. anthropic caches 5 minutes at 1.25x, or 1 hour at 2x. openai on gpt 5.6 and later keeps a fixed 30 minutes at 1.25x.
Running AI agents 24/7 on my side projects for a few weeks. What I learned:
- Cap the budget before it caps you
- Hourly checks burn more than real work
- Approve every public action
- Kill routines the day they stop paying off
Cut my agent costs a lot. What would you add?
One engineering lesson I'm seeing everywhere with agents
Use the expensive model less.
A lot of agent work is
- classify
- route
- choose a tool
- check a condition
- extract a value
- decide between 3 actions
You don't need frontier reasoning for every one of those steps.
I'd rather build
small model → routine decisions
frontier model → ambiguity + hard reasoning
deterministic code → everything that doesn't need AI
The best agent architecture might be the one that calls the smartest model the least.
Not every Hermes cron job needs an LLM.
For deterministic work, use no-agent mode:
Schedule → script → stdout
No model call.
No tokens.
No agent reasoning.
Great for:
• health checks
• data collection
• change detection
• simple recurring scripts
Save the LLM for tasks that actually need one.
I can confirm Sol 6.1high doesn't drain usage like Sol 6.0 or Astra. And it's great; the only downside is that it's slow
I have been running 4 orchestrators with multiple agents since Tuesday, and I’m still at 52%.
It would be long gone with 6 or Astra.
i'd put the token bill next to the ticket, not in a tooling bucket.
ticket | model | tokens | retries
that's the ledger i'd use to see where runs get expensive.
zap: claude code/codex, existing plugins/skills, open models. ~80% lower token cost.
One setting in Codex can make the same model burn 2.5× the credits of your GPT astra.
That’s GPT-6 Astra with Fast mode on.
Before buying more credits, check these 6 things
OpenAI has already documented:
1. Save Fast mode for work that needs it.
GPT-6 Astra consumes credits at 2.5× the Standard rate with Fast mode enabled.
If the task can wait, use Standard. In the CLI, run /fast off. Same model, lower credit rate. Official docs
2. Match the model to the task.
Try GPT-5.6 Luna for focused, routine tasks or Terra for everyday coding. Use Astra when the problem needs its capabilities.
Compare how much you spend getting a correct result. A smaller model helps when it can finish the job reliably.
3. Trim your AGENTS.md.
Keep shared instructions concise. Put folder-specific rules in nested AGENTS.md files.
Give Codex the relevant files, reproduction steps, and constraints. Unrelated documents and giant logs also consume tokens.
4. Disable MCP servers you aren’t using.
OpenAI explicitly lists MCP servers as extra context that consumes your allowance.
Keep the tools the task needs. Re-enable the others when they become useful.
Smaller models, concise instructions, and fewer unused MCP servers are all recommendations in OpenAI’s pricing docs.
5. Define what “done” means.
Name the behavior you want, the constraints, and how to verify it.
“Fix this bug, reproduce the original failure, and confirm the fix” gives Codex a clear finish line.
Separate required work from optional cleanup so a small fix doesn’t quietly become a renovation.
6. Give testing a stopping point.
OpenAI says Astra can test more broadly than small changes require.
Run the relevant tests and required checks. Once they pass, expand testing when new changes, failures, or unresolved concerns justify it.
A prompt to try:
“Complete the requested change within scope. Make reasonable assumptions for routine decisions and keep going. Run relevant tests and required checks. Stop when the acceptance criteria pass and no material concerns remain. Summarize the result briefly.”
Before upgrading your plan, see how much more useful work you can get from the allowance you already have.
Before scaling an AI batch job, look at the work you can avoid. Filter with ordinary data rules where possible, plan the remaining model calls, and measure quality alongside cost. A faster model is useful. Sending it thousands of unnecessary rows is still our design decision.
if you implement it yourself - keep in mind to only invoke ONE 1️⃣ subagent at the time
Fable LOOOVES spawning 20 simultaneously and it just kills your limit 🤯🔫
this setup allows me to run multiple Fable 5.1 High simultaneously, so try it out! 😇
Interesting finding while experimenting with OpenAI’s decision model for routing tasks in my workflow:
- Simple requests’ routing stage improved from roughly 1.3–1.6 seconds to 0.3–0.5 seconds.
- Across all 17 cases, estimated routing cost fell about 8%, but median routing latency increased about 13%.
Last week, I burnt about 60% of my Codex usage on one prompt w/ the newest model.
I wasn't going to let that happen again, so I asked Astra to fix my Codex setup for me.
It rebuilt how Codex chooses models:
Terra Medium = normal coding
Luna = routine/repetitive work
Sol = harder debugging, architecture, and review
Astra = only when the problem is hard enough to justify it
Now Codex can route work to the right model and escalate when needed instead of burning Astra on everything.
My goal wasn't to stop using Astra but was to stop wasting Astra.